Information-aware Message Passing Neural Networks for Graph Node Classification

In this paper, we present an information-aware message passing neural network (MPNN) to improve the effectiveness of weight calculation in the message passing process for node classification problems on a graph. Related researches are theoretically analyzed to introduce a special information interchanging scheme with precise weight generation. Compared to fix weighted message passing in convolutional networks (GCN) and neural network computed message coefficient in graph attention networks (GAT), the proposed information-aware message passing neural network takes advantage of node feature and its hidden representations, which indicate the amount of information should be passed and how to absorb the message between node neighbours. This exact message-passing method with self-learning for weight on each dimension of feature achieves better accuracy for node classification tasks in graph neural networks. The proposed method is evaluated by semi-supervised graph node classification task on citation network dataset. Obtained results show the correctness of the proposed method which are more accurate than that of GCN and GAT result.